On the regularization of Wasserstein GANs

Authors: Henning Petzka, Asja Fischer, Denis Lukovnikov

ICLR 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We perform several experiments on three toy data sets, 8Gaussians, 25Gaussians, and Swiss Roll, to compare the effect of different regularization terms. More specifically, we compare the performance of WGAN-GP and WGAN-LP as described in Equations (6) and (8) respectively... To see whether our findings on toy data sets can be transferred to real world settings, we trained bigger WGAN-GPs and WGAN-LPs on CIFAR-10...
Researcher Affiliation Collaboration Henning Petzka Fraunhofer Institute IAIS, Sankt Augustin, Germany henning.petzka@gmail.com Asja Fischer & Denis Lukovnikov Department of Computer Science, University of Bonn, Germany asja.fischer@gmail.com lukovnik@cs.uni-bonn.de
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks (clearly labeled algorithm sections or code-like formatted procedures).
Open Source Code Yes Code for the reproduction of our results is available under https://github.com/lukovnikov/improved_wgan_training .
Open Datasets Yes We perform several experiments on three toy data sets, 8Gaussians, 25Gaussians, and Swiss Roll, to compare the effect of different regularization terms...To see whether our findings on toy data sets can be transferred to real world settings, we trained bigger WGAN-GPs and WGAN-LPs on CIFAR-10 as it is described below.
Dataset Splits No The paper mentions 'validation loss' and 'validation set' in figures and text (e.g., 'Evolution of validation loss on CIFAR', Figure 16), implying its use. However, it does not explicitly provide specific dataset split percentages, sample counts, or refer to predefined validation splits needed to reproduce the data partitioning.
Hardware Specification No The paper does not explicitly describe the hardware used to run its experiments, lacking specific GPU/CPU models, processor types, or memory details.
Software Dependencies No The paper mentions using 'RMSprop' but does not provide specific version numbers for any software components, programming languages, or libraries used for the experiments.
Experiment Setup Yes Both, the generator network and the critic network, are simple feed-forward NNs with three hidden Leaky Re LU layers, each containing 512 neurons, and one linear output layer. The dimensionality of the latent variables of the generator network was set to two. During training, 10 critic updates are performed for every generator update, except for the first 25 generator updates, where the critic is updated 100 times for each generator update in order to get closer to the optimal critic in the beginning of training. Both networks were trained using RMSprop (Tijmen & Hinton, 2012) with learning rate 5 10 5 and a batch size of 256.